inference/research
verifiedmedium · 0.80updated 2026-10-15

tt-metal README featured-models performance table (Tenstorrent vendor, 2025-2026 releases)

benchmark

Fields

id
sources/tt-metal-readme-model-matrix-2026
Publisher
Tenstorrent
Kind
benchmark
Accessed
2026-10-15
notes
Vendor model-matrix table: TTFT(ms), T/S/U (tok/s per user = 1/inter-token-latency), T/S (= T/S/U x batch) on tt-metal model demos. ISL=128 for all reported LLM rows. Hardware tags: Galaxy (Wormhole, TP32), QuietBox (Wormhole, TP8 = 8x n300-class chips), n300 (TP2). TT-Metalium releases v0.62-v0.65 pinned per row. Vendor-measured on their own demos.

Notes

Vendor model-matrix table: TTFT(ms), T/S/U (tok/s per user = 1/inter-token-latency), T/S (= T/S/U x batch) on tt-metal model demos. ISL=128 for all reported LLM rows. Hardware tags: Galaxy (Wormhole, TP32), QuietBox (Wormhole, TP8 = 8x n300-class chips), n300 (TP2). TT-Metalium releases v0.62-v0.65 pinned per row. Vendor-measured on their own demos.

Cross-references

This record references no other records.